Decomposition of complex numerical software into cooperating components

Abstract

This paper discusses the approach that is being taken for parallelising the WAQUA/Kalman software of the Dutch Institute for Coastal and Marine Management. This software is used for numerical simulation of flow and transport phenomena in coastal waters and incorporates a Kalman filtering procedure to assimilate observational data into the simulation. In particular the Kalman filtering part of the software is very time-consuming which prohibits its use in e.g. operational storm surge prediction. Parallelisation is being considered to reduce this run-time.

Instead of parallelising the code as it is, it is first decomposed into a number of components which are coupled through an advanced message passing library. After this, each of the components can be parallelised in the way that is most appropriate for the computations in the component. If the original code were parallelised directly, it would not be possible to accommodate the various forms of parallelism that are available in the program. The decomposition into components itself already gives a clear benefit because it enhances the maintainability and reusability of the code.